docs: S2 learning-curve design v0 (proposal, no run)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
ClaudeandClaude Opus 5.5 committed 2026-11-14 08:54:10 +00:00
1 parent 628a10bad3
commit d8aa0f4206
1 file changed
+65
+65
View File
@@ -0,0 +1,65 @@
# S2 design v0: learning-curve test on Greif v5 (Claude, 2026-11-14)
Status: proposal for review. Nothing runs before Yanqi gives an explicit go, because S2 is the first stage that trains models.
Inputs: S1 result (codex/s1-real-scoring 90b8c66), my S1 review, and MoldAstra's note 13 (Drive 1uhaP8gEd15_g5xnFcX7kIeI3mJwCfo3o).
## 1. Question
Do edge-direction readings (O2 edge q mean/max) lower the arrival-time error on *unseen source groups*, or the number of training sources needed to reach a given error, when added to a strong whole-part model with cost-accounted distance features?
Not asked in this round:
- O2 node invariants: S1 gave no coverage support; they may return as a later arm.
- O1 partitioning: its type definition and region tables are still open, and it is not refuted.
## 2. Arms (one factor; identical backbone, data, budget and seeds)
- **R2 baseline:** geodesic/eikonal-from-gate features plus per-node geometric scalars, with GBRT (Gre24 style). Feature cost is itemised.
- **N (control):** a whole-part neural backbone with base inputs:
- normalised coordinates plus physical scale (R_g, volume);
- d/t_part;
- gate distance / R_g;
- geodesic distance;
- **g12 and g23**.
- **E:** N plus the 4 edge-direction readings (mean/max of q1 and q2 over true tet edges).
N and E carry the same g12 and g23, so a gain can be attributed to the direction readings and not to the spectral gaps (MoldAstra note 13).
**Placement of the local features:** the backbone encodes the whole part from the base inputs at all nodes. O2 features (g, and edge q in arm E) enter only at *query nodes*, the sampled anchors, through the decoder. This keeps O2 cost at sampled nodes, as in S1, and isolates their contribution.
## 3. Data, split, targets
- **Corpus:** all 624 Greif v5 parts.
- **Prerequisite:** a near-duplicate rule across source groups (MoldAstra), giving grouping key = source key union near-duplicate cluster.
- **Prerequisite:** d/t_part on every part (counted cost: the panel took 953 s for 24 parts, so the full corpus is about 7 CPU-h before parallelism).
- **Split:** 5-fold grouped CV over the final groups. Training pools are nested at 25/50/100 % of the training groups, each trained from scratch. 3 seeds per point.
- **Query sampling:** 2000 mass-PPS draws per part with replacement, the S1 design and seed rule. The same draws serve as training targets and evaluation queries.
- O2 cost: the panel needed 6.6e9 visits for 24 parts, so roughly 1.7e11 for 624, about 19 CPU-h at 0.4 us/visit. Counting only per part first; 30 min / 1.25 GiB per field.
- **Target:** Fill_time at query nodes, normalised per part by its maximum fill time.
- **Metrics:**
- Primary: the volume-weighted relative L2 error per part (ratio estimator over draws), source-equal mean across test groups.
- Secondary:
- last-filled-region location error;
- Kendall tau on sampled node pairs;
- near-wall slice;
- worst group.
## 4. Decision (pre-registered)
- Δ = relative error reduction of E vs N. Positive Δ means E has lower error.
- The interval [L, U] is a group bootstrap over test groups, with seeds nested.
- Margin m = 10 % (candidate; frozen before the first run).
- The six-region table from the design record §6 applies.
- R2 is reported in the same table. N and E must beat R2 for the neural route to be called useful at all.
## 5. Compute
- **Pilot first:** one fold at the 100 % pool, measuring t_step and peak memory on one V100 16 GB with FP16 AMP.
- **Matrix size:** 3 arms x 3 pools x 3 seeds x 5 folds = 135 runs, with R2 on CPU.
- Budget from the measured pilot, using the formula GPU-h = Σ epochs x steps x t_step x (1 + overhead).
- Backbone and size are chosen at the pilot, not here.
## 6. Deferred, with reasons
- **MoldAstra's bypass pair:** same downstream slit under constant Δp, upstream bypass open vs closed. Analytic T_closed − T_open = A·R·L / (2Δp). It tests whether a model uses connectivity, and generic vs type partitions.
- It needs paired reference solves, so it belongs to the in-house CAE stage and is not an S2 start condition.
- Under fixed flow with a single downstream outlet the two cases arrive at the same time. Both control modes must therefore be included.
- **O1 arms:** after the O1 type definition and admission rule are frozen.
## 7. Open before any run
1. Near-duplicate rule (MoldAstra) and the frozen grouping.
2. Freeze m and the S2 seed table.
3. Yanqi's explicit go for training.